The future of work: How AI is transforming remote work

My entire job takes place on my laptop.

I write stories like this in Google Docs on my laptop. I coordinate with my editor in Slack on my laptop. I reach out to sources with Gmail and then interview them over Zoom, on my laptop. This isn’t true of all journalists — some go to war zones — but it’s true of many of us, and for accountants, tax preparers, software engineers, and many more workers, maybe over one in 10, besides.

Laptop jobs have many charms: the lack of a commute or dress code, the location flexibility, the absence of real physical strain. But if you’re a laptop worker and not worried about what’s coming in the next decade, you haven’t been paying attention. There is no segment of the labor market more at risk from rapid improvements in AI than us.

The newest “reasoning models” from top AI companies are already essentially human-level, if not superhuman, at many programming tasks, which in turn has already led new tech startups to hire fewer workers. Generative AIs like Dall-E, Sora, or Midjourney are actively competing with human visual artists; they’ve already noticeably reduced demand for freelance graphic design.

Services like OpenAI’s Deep Research are very good at internet-based research projects like, say, digging up background information for a Vox piece. “Agentic” AIs like Operator are able to coordinate and sequence these kinds of tasks the way a good manager might. And the rapid pace of progress in the field means that laptop warriors can’t even take comfort in the fact that current versions of these programs and models may be janky and buggy. They will only get better from here, while we humans will stay mostly the same.

As AIs have improved at laptop job tasks, progress on more physical work has been slower. Humanoid robots capable of tasks like folding laundry have been a longtime dream, but the state-of-the-art falls wildly short of human level. Self-driving cars have seen considerable progress, but the dream has proven harder to achieve than boosters thought. While AI has been improving rapidly, robotics — the ability of AI to work in the physical world — has been improving much more slowly. At this point, a robot plumber or maid is far harder to imagine than a robot accountant or lawyer.

Let me offer, then, a thought experiment. Imagine we get to a point — maybe in the next couple years, maybe in 10, maybe in 20 — when AI models can fully substitute for any remote worker. They can write this article better than me, make YouTube videos more popular than Mr. Beast’s, do the work of an army of accountants, and review millions of discovery documents for a multibillion-dollar lawsuit, all in a matter of minutes. We would have, to borrow a phrase from AI writer and investor Leopold Aschenbrenner, “drop-in remote workers.” How does that reshape the US, and world, economy?

Right now this is a hypothetical. But it’s a hypothetical worth taking seriously — seriously enough that I may or may not be visiting the International Brotherhood of Electrical Workers’ apprenticeship application most days, just in case I need work that requires a human body.

Fast AI progress, slow robotics progress

If you’ve heard of OpenAI, you’ve heard of its language models: GPTs 1, 2, 3, 3.5, 4, and most recently 4.5. You might have heard of their image generation model DALL-E or video generation model Sora.

But you probably haven’t heard of their Rubik’s cube solving robot. That’s because the team that built it was disbanded in 2021, about a year before the release of ChatGPT and the company’s explosion into public consciousness.

OpenAI engineer Wojciech Zaremba explained on a podcast that year that the company had determined there was not enough real-world data of how to move in the real world to keep making progress on the robot. Two years of work, between 2017 and 2019, was enough to get the robot hand to a point where it could unscramble Rubik’s Cubes successfully 20 to 60 percent of the time, depending on how well-scrambled the Cube was. That’s … not especially great, particularly when held up next to OpenAI’s language models, which even in earlier versions seemed capable of competing with humans on certain tasks.

It’s a small story that encapsulates a truism in the AI world: the physical is lagging the cognitive. Or, more simply, the chatbots are beating the robots.

This is not a new observation: It’s called Moravec’s paradox, after the futurist Hans Moravec, who famously observed that computers tend to do poorly at tasks that are easy for humans and do well at tasks that are often hard for humans.

Why? Here we’re less sure. As the machine learning researcher Nathan Lambert once noted, Moravec’s paradox is “based on observation, not theory. We have a lot of work to do to figure out why.” But we have some hypotheses.

Perhaps human-like motions are harder for robots because we gained them relatively early in evolutionary time, far earlier than our capacity for reasoning. Running on two or even four legs is a very old ability that many animals share; it’s instinctual for us, which both makes it harder for machines without that evolutionary history to learn, and harder for us to articulate to those machines.

Harder still is the fact that a robot has to learn to run on two legs by actually running on two legs in real life. This point is key: If OpenAI had its servers pronouncing every sentence that ChatGPT generates, out loud, one at a time, as part of its training process, it probably would’ve taken millennia to get to today’s abilities. Instead, it was able to train the GPT models using millions of CPU cores operating in parallel to analyze vast reams of data, processing trillions of individual words a second. Each new model only requires months or a few years of training because the process happens much, much faster than real time.

Historically roboticists’ way around this limitation was to make simulated worlds, sort of purpose-built video game environments, in which to train robots much faster. But when you take the bot out of the virtual playground and into the real world, it has a tendency to fail. Roboticists call this the “sim2real” (simulation to reality) gap, and many a noble robot has fallen into it (and over it, and on it) over the years.

The optimistic theory of the case is that, given enough real-world data about movement, the same techniques that have made language models so successful can be used to make robots work well. The most bullish takes on robotics I’ve seen, like this from Anthropic co-founder Jack Clark last year, are based on the idea that if you throw enough data (from stuff like YouTube videos of people walking around, or from actual humans operating the robot with a controller) into well-enough designed and fine-tuned transformer models (using the same learning structure as ChatGPT or Claude etc.), the end result will be a model good enough to govern a robot in the real world.

Maybe! So far we mostly have academic demonstrations rather than the real-world, commercialized products that large language models are today. (Disclosure: Vox Media is one of several publishers that has signed partnership agreements with OpenAI. One of Anthropic’s early investors is James McClave, whose BEMC Foundation helps fund Future Perfect. Our reporting remains editorially independent.)

I don’t know the trajectory of cognitive AI and robotics over the next decade. Maybe, as OpenAI CEO Sam Altman has predicted, this year will “see the first AI agents ‘join the workforce’ and materially change the output of companies” (referring, presumably, to software workers rather than robots). Maybe, as critics argue, the cost of training these models will prove too immense and the companies developing them, which are burning through billions in VC funding, will fail. Maybe robotics will continue to lag, or maybe people will have Rosie from The Jetsons dusting their furniture next year. I have my guesses, but I know enough to know I shouldn’t be too confident.

My median guess, though, is the world outlined above: language, audiovisual, and otherwise non-physical models continue to make very rapid progress, perhaps becoming capable of doing any fully remote job currently done by humans within the next decade; robotics continues to lag, being very useful in advanced manufacturing but unable to garden or change your sheets or empty your dishwasher. Taken to an extreme, this could look like, in the words of Anthropic CEO Dario Amodei, a “country of geniuses in a datacenter.”

What does that world look like?

The work left for the rest of us

One of the more useful pieces examining this idea came out in January from Epoch AI, a small research group that’s quickly become the most reliable source of data on cutting-edge AI models. The author, Matthew Barnett, uses a commercially available AI model (GPT-4o) to go through a US Department of Labor-sponsored database of over 19,000 job tasks and categorize each of them as doable remotely (writing code, sending emails) or not doable remotely (firefighting, bowling).

A task, notably, is not the same as a job or occupation. The occupation “journalist” includes specific subtasks like writing emails, composing articles, making phone calls, appearing on panels, reading academic papers, and so on. And an occupation as a whole cannot be automated unless all tasks, or at least all absolutely necessary tasks, can themselves be automated. An AI might be able to do some of the mental labor a surgeon has to perform, for instance, but until it can actually cut and suture a human, the surgeon’s job remains safe.

Barnett finds that 34 percent of tasks can be performed remotely, but only 13 percent of occupations have, as their top five most important subtasks, things that can all be done remotely. Thirteen percent can then serve as an (admittedly very rough) estimate of the share of jobs that could, in principle, be fully automated by a sufficiently advanced cognitive AI.

Obviously, a world in which 13 percent of jobs are rapidly automated away is one with pretty massive social disruption. But at first glance, it doesn’t seem too different from what’s been happening in many industries over the past couple of centuries. In 1870, about half of United States workers worked in agriculture. By 1900, a third did. Last year, only 1.4 percent did. The consequence of this is not that Americans starve, but that a vastly more productive, heavily automated farming sector feeds us and lets the other 98.6 percent of the workforce do other work we like more.

Similarly, manufacturing has become so automated that it now appears global manufacturing employment has peaked — it’s not just that factories use fewer workers in the US compared to poorer countries, but that they use fewer workers everywhere, period.

“There’s an upper bound of how much can be remote, and I think we’re kind of at it now.”

— Nicholas Bloom, Stanford University economist and leading expert on remote work

Agriculture and manufacturing are also becoming less important as a share of global economic output over time, not just as shares of employment. So this is one possible future: AI rapidly increases productivity in remote-friendly jobs like software engineering, accounting, and writing for Vox.com, leading to sharp reductions in employment in those sectors. People displaced by this shift gradually shift to harder to automate jobs, becoming masseuses, electricians, nurses, and so forth.

Barnett notes that if this happens, the effect on global economic growth could be massive (maybe a doubling of economic output). It would obviously be inconvenient for me, personally, and I would be sad. But it’s basically “the world now, but moreso” — more economic growth and more labor displacement — rather than a brave new world.

That said Barnett thinks this is probably underselling what might happen. Yes, automation in agriculture and manufacturing has meant that those sectors gradually decline in importance. That doesn’t have to happen, though. Barnett gives the example of the UK after the invention of spinning jenny and flying shuttle. Those and subsequent cotton-processing technologies massively improved productivity in the textiles industry relative to other sectors of the British economy.

Was the result that textiles became less important? Quite the opposite: The sector exploded, and became vastly more important to the British economy. British exports of textiles increased over sevenfold between the 1740s (when those inventions were just being developed and deployed) and the 1750s, and kept growing exponentially from there.

The difference between these scenarios is a number that Barnett calls the “elasticity of substitution” — in this case, between remote and in-person work, but in principle between any two kinds of work. For some kinds of work, this number is below 1, meaning that if that work gets much cheaper, it will shrink relative to other kinds of work. The two types of work don’t substitute well for each other, so the elasticity of substitution is low. But if the elasticity is above 1, then the work getting cheaper will become more common and more important.

One way to think about this, Barnett told me, whether your demand for something can be saturated. “There’s a sense in which your utility from food saturates, because the amount of utility you get from just getting 2,000 calories per day is not half the amount of utility you get from 4,000.” he told me. “Assuming you can live comfortably on 2,000 calories per day, then it’s going to be almost exactly the same amount of utility, because you’re probably gonna throw away a whole bunch of food.”

It makes sense, then, that agriculture shrank in importance once humanity developed the ability to grow more calories than people needed (the world’s farms currently produce about 3,000 calories per person per day, more than enough to feed every human on the planet by sheer quantity). Manufacturing, too, makes some sense in these terms. Most people hit a limit on how much large manufactured stuff they actually are able to use. My first washing machine helped a lot; getting a third or even a second would be pointless.

By contrast, the world’s demand for textiles in the 18th century was nowhere near a saturation point. You can, in principle, own a limitless supply of clothes, and especially in the time of hand production, there was lots of pent-up demand, in countries around the world, for fabrics that had previously been prohibitively expensive. That meant that Britain could pour more and more resources into that sector of its economy without having returns diminish too much.

What if remote work is more like that?

This supposition might seem fanciful, but let’s fantasize. If you had an on-call computer programmer who could make your computer work exactly the way you wanted, wouldn’t you have a lot to ask it? If you had a personal animator who could make on-demand episodes of your favorite type of TV show with your favorite music in the background, wouldn’t you call on her a lot?

I have a million deeply weird questions I’m too busy and/or lazy to answer — who invented the “You Can’t Hurry Love” bassline? Why were the witness reports in the Dag Hammarskjold plane crash ignored? — that I wish something smarter than OpenAI Deep Research could give me an answer in seconds. Maybe you would too?

If that’s the situation, then things look very different. If the elasticity of substitution between remote and non-remote work is 10, Barnett finds, then you see US GDP grow tenfold over a decade, an average growth rate of 25 percent. That is completely unheard of in human history. But it would also be incredibly weird growth, showing up in increased consumption of AI-generated products, rather than, say, easier access to child care or cheaper housing.

Nicholas Bloom, the Stanford University economist and leading expert on remote work, is taking the under on this bet. It’s better, he reasons, to think of remote and non-remote work as complements than substitutes, which makes a scenario with high substitution, like Barnett’s fast growth situation, hard to believe.

“There’s an upper bound of how much can be remote, and I think we’re kind of at it now,” Bloom says. That said, part of Bloom’s skepticism about full-remote work comes from his belief in the importance of mentoring, which is much harder to do in a remote work setup. With AI, presumably the need to mentor in-person becomes moot.

What are the most remote-friendly jobs?

One can of course reason through which jobs are easy to do remotely (graphic design, telemarketing) and which are impossible (surgery, construction). But is it possible to be more systematic?

Several researchers have tried to categorize major occupations as remote-able or not, but I like Matthew Barnett’s approach of simply asking a large language model if certain tasks can be done remotely. Here are some examples of jobs where every single task can be done remotely, per the OpenAI model that Barnett asked (GPT-4o):

Bioinformatics scientists
Bioinformatics technicians
Business continuity planners
Business intelligence analysts
Clinical data managers
Credit analysts
Credit counselors
Customer service representatives
Data warehousing specialists
Database administrators
Database architects
Editors
Environmental economists
Financial quantitative analysts
Geographic information systems technologists and technicians
Information security analysts
Information technology project managers
Insurance underwriters
Mathematicians
Preventive medicine physicians
Proofreaders and copy markers
Search marketing strategists
Securities, commodities, and financial services sales agents
Telemarketers
Travel agents
Video game designers
Web administrators
Web developers
Writers and authors

How doomed are remote workers?

Before getting too carried away, it’s worth remembering — we’re not here, yet. At the very least, an AI remote worker will have to use a computer fluently, and perhaps surprisingly, the best benchmarks we have, like OSWorld, do not show AI models doing that. “The fact is right now that models really suck at navigating browsers,” Jaime Sevilla, who runs the Epoch forecasting group, told me. “They’re not at the level of my grandmother currently.”

Sevilla suggested that the pace of investment and progress he’s seeing suggests that we might get grandma-level Chrome usage within a year or two. But it’ll be some time from there to actually using Chrome in an economically useful way, or managing a developer team in Slack, or any number of other specific tasks we expect remote workers to do.

We’ll also probably learn a great deal about the character of the jobs we’re automating. Tamay Besiroglu, also at Epoch, notes that AI became superhuman at playing chess in 1997, when IBM’s Deep Blue defeated world champion Garry Kasparov. Today the top chess engine, Stockfish, is wildly, wildly better than the top-ranked human player, Magnus Carlsen. But chess is arguably more popular than it’s ever been. Carlsen is a global celebrity with more than 1.4 million subscribers on his YouTube channel, where he streams matches and analyzes games from shows like The Queen’s Gambit. His job has been automated to hell, and he’s a millionaire.

“We discovered that, actually, the thing that people pay chess players for isn’t their ability to produce very good chess moves,” Besiroglu concludes. “That’s one thing, but not the entire thing. Things like being entertaining, having charisma, being a good streamer — those things are very important. And we don’t have good benchmarks for how entertaining or charismatic an AI system is.”

To be fair, Besiroglu expects AI to gain those capabilities in the next five to 10 years. But even if it does, I think it’s plausible that people will still be willing to pay for a connection to a specific human, a connection that AI, by its very nature, cannot fully replace. Magnes Carlsen the chess player can be, and has been automated; it’s less obvious to me that Magnes Carlsen, the influencer, can be automated as well.

So I’m not hanging up my laptop and giving up just yet. Maybe people will still value human-grown hot takes, the way they value Magnus Carlsen’s human-developed chess style. Or maybe not, in which case, electrician school might start looking better.

But I keep thinking back to Barnett’s conclusion that human-level cognitive AI could maybe do 13 percent of work out of the box. A world where those are the only jobs that get automated is not a situation where the singularity happens (that is, where AI becomes so good that it is capable of recursively improving itself without human intervention and eventually becomes superhuman in all tasks). But it’s one where society is transformed radically all the same.

When I talk to people working in AI, they treat that transformation as nearly inevitable, perhaps a lowball for the changes that might actually be on their way. When I talk to everyone else, I get the sense they have no idea what’s coming.